Research map / Detection and segmentation / CNN features
Looking Locally: Object-Centric Vision Transformers as Foundation Models for Efficient Segmentation
Manuel Traub, Martin V. Butz
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use fovea-like input patching with FLIP-Large instead of full-image SAM2-L | FLIP · Small objects (<1% image area), average | mean IoU (%): 73.50 -> 81.19 | Improved |
| use FLIP-Large instead of SAM-H | SAM-H · Average over 6 datasets (incl. ObjaScale) | mean IoU (%): 75.41 -> 83.26 | Improved |
| use FLIP-Tiny (0.51M parameters) instead of SAM2-L | SAM2-L · Average over 6 datasets (incl. ObjaScale) | mean IoU (%): 75.87 -> 79.90 | Improved |
| use FLIP-Large instead of SAM2-L | SAM2-L · Inference time (RTX4090) | time (ms): 88.64 -> 44.13 | Improved |
| scale FLIP from Tiny to Large | FLIP · Average over 6 datasets (incl. ObjaScale) | mean IoU (%): 79.90 -> 83.26 | Improved |
| scale FLIP from Tiny to Large | FLIP · KITTI-360 | mean IoU (%): 72.62 -> 70.72 | Hurt |
| scale FLIP_bb from Tiny to Large | FLIP_bb · Box prompt noise 20% | mean IoU (%): 30.76 -> 26.34 | Hurt |
| use box-derived Gaussian prompts (FLIP_bb-Large) instead of SAM-H under 20% box noise | SAM-H · Box prompt noise 20% | mean IoU (%): 47.43 -> 26.34 | Hurt |
| use FLIP_bb-Large with clean boxes instead of SAM-H | SAM-H · Box prompt noise clean (subset) | mean IoU (%): 75.36 -> 79.53 | Improved |
| use native Gaussian prompts (FLIP-Large) instead of box-derived (FLIP_bb-Large) (ablation) | FLIP_bb-Large · Average over 6 datasets (incl. ObjaScale) | mean IoU (%): 79.46 -> 83.26 | Improved |
| use FLIP-Tiny instead of SAM-H on ObjaScale | SAM-H · ObjaScale | mean IoU (%): 73.76 -> 88.55 | Improved |
| use SAM2-L instead of SAM-H on ObjaScale | SAM-H · ObjaScale | mean IoU (%): 73.76 -> 69.54 | Hurt |
Closest papers by meaning
- VGGT-Segmentor: Geometry-Enhanced Cross-View Segmentation (CVPR 2026 · similarity 0.87)
- Segment This Thing: Foveated Tokenization for Efficient Point-Prompted Segmentation (CVPR 2025 · similarity 0.87)
- FocSAM: Delving Deeply into Focused Objects in Segmenting Anything (CVPR 2024 · similarity 0.86)
- M^3-VOS: Multi-Phase, Multi-Transition, and Multi-Scenery Video Object Segmentation (CVPR 2025 · similarity 0.86)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.85)
- ODIN: A Single Model for 2D and 3D Segmentation (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
